US20240086707A1 - Device, method, program, and system for detecting unidentified water - Google Patents
Device, method, program, and system for detecting unidentified water Download PDFInfo
- Publication number
- US20240086707A1 US20240086707A1 US18/516,032 US202318516032A US2024086707A1 US 20240086707 A1 US20240086707 A1 US 20240086707A1 US 202318516032 A US202318516032 A US 202318516032A US 2024086707 A1 US2024086707 A1 US 2024086707A1
- Authority
- US
- United States
- Prior art keywords
- water
- acoustic
- unknown
- acoustic data
- feature amount
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 355
- 238000000034 method Methods 0.000 title claims description 35
- 238000001514 detection method Methods 0.000 claims abstract description 112
- 238000010801 machine learning Methods 0.000 claims abstract description 18
- 230000002123 temporal effect Effects 0.000 claims abstract description 18
- 239000000284 extract Substances 0.000 claims abstract description 11
- 230000008569 process Effects 0.000 claims description 20
- 238000013528 artificial neural network Methods 0.000 claims description 19
- 230000008859 change Effects 0.000 claims description 18
- 239000002351 wastewater Substances 0.000 claims description 12
- 238000004891 communication Methods 0.000 claims description 10
- 239000003673 groundwater Substances 0.000 claims description 8
- 230000015654 memory Effects 0.000 claims description 6
- 235000021162 brunch Nutrition 0.000 claims 4
- 238000000605 extraction Methods 0.000 abstract description 22
- 239000013598 vector Substances 0.000 description 20
- 230000003595 spectral effect Effects 0.000 description 11
- 239000011159 matrix material Substances 0.000 description 6
- 239000010865 sewage Substances 0.000 description 6
- 238000012545 processing Methods 0.000 description 5
- 230000011218 segmentation Effects 0.000 description 5
- 238000012549 training Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 238000009434 installation Methods 0.000 description 3
- 238000007781 pre-processing Methods 0.000 description 3
- 238000004065 wastewater treatment Methods 0.000 description 3
- 230000004907 flux Effects 0.000 description 2
- 230000003936 working memory Effects 0.000 description 2
- 230000002547 anomalous effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 235000021152 breakfast Nutrition 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- E—FIXED CONSTRUCTIONS
- E03—WATER SUPPLY; SEWERAGE
- E03F—SEWERS; CESSPOOLS
- E03F1/00—Methods, systems, or installations for draining-off sewage or storm water
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01F—MEASURING VOLUME, VOLUME FLOW, MASS FLOW OR LIQUID LEVEL; METERING BY VOLUME
- G01F1/00—Measuring the volume flow or mass flow of fluid or fluent solid material wherein the fluid passes through a meter in a continuous flow
- G01F1/66—Measuring the volume flow or mass flow of fluid or fluent solid material wherein the fluid passes through a meter in a continuous flow by measuring frequency, phase shift or propagation time of electromagnetic or other waves, e.g. using ultrasonic flowmeters
- G01F1/666—Measuring the volume flow or mass flow of fluid or fluent solid material wherein the fluid passes through a meter in a continuous flow by measuring frequency, phase shift or propagation time of electromagnetic or other waves, e.g. using ultrasonic flowmeters by detecting noise and sounds generated by the flowing fluid
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L25/00—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
- G10L25/48—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
- G10L25/51—Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01F—MEASURING VOLUME, VOLUME FLOW, MASS FLOW OR LIQUID LEVEL; METERING BY VOLUME
- G01F15/00—Details of, or accessories for, apparatus of groups G01F1/00 - G01F13/00 insofar as such details or appliances are not adapted to particular types of such apparatus
- G01F15/06—Indicating or recording devices
- G01F15/061—Indicating or recording devices for remote indication
- G01F15/063—Indicating or recording devices for remote indication using electrical means
Definitions
- the present invention relates to a technique for detecting unknown water using a machine learning model.
- a separate sewer system is currently in the mainstream of sewage treatment, in which a wastewater pipeline for wastewater discharged from houses, factories, and the like and a rainwater pipeline for rainwater are separately installed.
- the wastewater discharged into the wastewater pipeline is treated in a sewage treatment plant and is let out into a river or the like after treatment. Meanwhile, the rainwater that has flowed into the rainwater pipeline is directly let out into a river or the like.
- the amount of wastewater flowing into the sewage treatment plant is regarded as a wastewater treatment amount for treatment at the sewage treatment plant.
- a sewerage fee corresponding to water supply usage is charged for the wastewater treatment amount (revenue water amount).
- the amount of water corresponding to this difference is called unknown water.
- the unknown water is rainwater, groundwater, and/or the like that somehow flow into the wastewater pipeline.
- the wastewater having flowed into the wastewater pipeline has to be treated in the sewage treatment plant.
- the unknown water places stress on management of sewage treatment business.
- the 2011 statistics indicate that the unknown water is 17.1% of the annual total amount of treated water.
- each of the flowmeters is costly in terms of not only the flowmeter itself, but also the work cost and labor cost for installation and/or the like, and therefore, a technique for detecting unknown water which can be implemented at lower cost has been required.
- An object of the present invention considering the above-mentioned problem is to provide a technique for achieving detection of unknown water without the need for a flowmeter.
- One aspect of the present invention relates to an unknown-water detection apparatus that, in order to solve the above problem, includes a feature amount extraction section that extracts, from acoustic data including a water flowing sound, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount; and an unknown-water prediction section that predicts presence or absence of unknown water from an acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned an acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time, in which the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot.
- detection of unknown water can be achieved without the need for a flowmeter.
- FIG. 1 illustrates an outline of an unknown-water detection apparatus according to one example of the present invention
- FIG. 2 is a block diagram illustrating a hardware configuration of the unknown-water detection apparatus according to one example of the present invention
- FIG. 3 is a block diagram illustrating a functional configuration of the unknown-water detection apparatus according to one example of the present invention
- FIG. 4 illustrates an acoustic feature amount pattern in a matrix format according to one example of the present invention
- FIG. 5 is a flowchart illustrating an unknown-water detection process according to one example of the present invention.
- FIG. 6 A schematically illustrates an acoustic feature amount pattern according to one example of the present invention
- FIG. 6 B schematically illustrates an acoustic feature amount pattern according to one example of the present invention
- FIGS. 7 A and 7 B illustrate a specific example of a temporal change in water usage and acoustic feature amount in one day
- FIG. 8 illustrates a prediction result on the presence or absence of unknown water according to one example of the present invention
- FIG. 9 schematically illustrates an unknown-water detection model using a subspace method according to one example of the present invention.
- FIG. 10 is a flowchart illustrating an unknown-water prediction process using the subspace method according to one example of the present invention.
- FIG. 11 is a flowchart illustrating a learning process for an unknown-water detection model implemented using a neural network according to one example of the present invention.
- FIG. 12 illustrates an unknown-water prediction result by the unknown-water detection apparatus according to one example of the present invention.
- an unknown-water detection apparatus that detects unknown water from acoustic data including a water flowing sound in a sewer pipe.
- the example described below is outlined as follows: when the unknown-water detection apparatus obtains acoustic data including a water flowing sound in a sewer pipe collected by a sound collection apparatus installed down in a manhole, the unknown-water detection apparatus extracts an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount from the obtained acoustic data, and predicts the presence or absence of unknown water from the extracted acoustic feature amount pattern by utilizing a learned unknown-water detection model.
- the unknown-water detection model may be a machine learning model that learned an acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time.
- the unknown-water detection model may be a machine learning model that uses a subspace to characterize the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time (i.e., an acoustic feature amount pattern of normally used water not including unknown water).
- the unknown-water detection model may be a neural network that learned the acoustic feature amount pattern extracted from the acoustic data during the no-rainfall time that does not include the water flowing sound of unknown water, and from the acoustic data obtained under the condition different from the no-rainfall time that includes the water flowing sound of unknown water.
- the unknown-water detection apparatus can detect the presence or absence of unknown water based on the acoustic data obtained by utilizing a sound collection apparatus such as a commercially available voice recorder, so that it becomes possible to perform unknown-water detection at lower cost than in unknown-water detection using the conventional flowmeter.
- FIG. 1 schematically illustrates the unknown-water detection apparatus according to one example of the present invention.
- unknown-water detection apparatus 100 when unknown-water detection apparatus 100 obtains acoustic data including a water flowing sound in a sewer pipe obtained by a sound collection apparatus such as a voice recorder installed in a manhole, the unknown-water detection apparatus utilizes an unknown-water detection model implemented as a machine learning model to predict the presence or absence of unknown water from the obtained acoustic data.
- a sound collection apparatus such as a voice recorder installed in a manhole
- the sound collection apparatus may be a commercially available voice recorder or the like. Sound collection apparatuses are installed down in a plurality of manholes in an area, and record, as acoustic data in a predetermined format or acoustic waveform, sounds (water flowing sounds) of water flowing through sewer pipes down in the manholes.
- the acoustic data recorded in each of the sound collection apparatuses may be appropriately collected by hands or the like or may also be collected via a communication network when the sound collection apparatus is provided with a communication function.
- Unknown-water detection apparatus 100 preprocesses the obtained acoustic data and segments the obtained acoustic data into segments of data of a predetermined period of time (e.g., 10 minutes) to generate an acoustic feature amount pattern indicating a temporal change of time-serialized acoustic data (e.g., in one day, one week, or the like). Unknown-water detection apparatus 100 inputs the generated acoustic feature amount pattern into the learned unknown-water detection model, to obtain a prediction result indicating the presence or absence of unknown water from the unknown-water detection model.
- a predetermined period of time e.g. 10 minutes
- Unknown-water detection apparatus 100 inputs the generated acoustic feature amount pattern into the learned unknown-water detection model, to obtain a prediction result indicating the presence or absence of unknown water from the unknown-water detection model.
- unknown-water detection apparatus 100 may have a hardware configuration including processor 101 such as a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), or the like, memory 102 such as a Random Access Memory (RAM), a flash memory, or the like, hard disk 103 , and input/output (I/O) interface 104 as illustrated in FIG. 2 , for example.
- processor 101 such as a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), or the like
- memory 102 such as a Random Access Memory (RAM), a flash memory, or the like
- hard disk 103 such as a hard disk 103
- I/O input/output
- Processor 101 executes various processes of unknown-water detection apparatus 100 , which will be described in detail below.
- Memory 102 stores therein various data and programs for unknown-water detection apparatus 100 together with a program for implementing the unknown-water detection model, and functions as a working memory especially for working data, a program being executed, and the like. Specifically, memory 102 stores therein data and programs loaded from hard disk 103 and functions as the working memory during program execution by processor 101 .
- Hard disk 103 stores therein the various data and programs for unknown-water detection apparatus 100 together with the program for implementing the unknown-water detection model.
- I/O interface 104 is an interface for inputting and outputting data into and from an external apparatus, and is a device for inputting and outputting data of a Universal Serial Bus (USB), a communication circuit, a keyboard, a mouse, a display, and the like, for example.
- USB Universal Serial Bus
- unknown-water detection apparatus 100 is not limited to the above-described configuration, and the unknown-water detection apparatus may have any other suitable hardware configuration.
- a part or all of the unknown-water detection process of unknown-water detection apparatus 100 described above may be implemented by a processing circuit or an electronic circuit wired to implement the unknown-water detection process.
- FIG. 3 is a block diagram illustrating a functional configuration of unknown-water detection apparatus 100 according to one example of the present invention. As illustrated in FIG. 3 , unknown-water detection apparatus 100 includes feature amount extraction section 110 and unknown-water prediction section 120 .
- Feature amount extraction section 110 extracts, from the acoustic data including the water flowing sound, the acoustic feature amount pattern indicating a temporal change in an acoustic feature amount.
- feature amount extraction section 110 first performs preprocessing and segmentation on the acoustic data indicating an acoustic waveform representing the water flowing sound in a sewer pipe.
- feature amount extraction section 110 first discards, when the acoustic data includes any data loss caused by a malfunction of the sound collection apparatus or the like, a zero value or non-numerical data in the lost part.
- feature amount extraction section 110 segments the obtained continuous acoustic waveform into segments at predetermined time intervals in order to facilitate subsequent data processing.
- the segmentation may divide the acoustic waveform data into data pieces having 10-minute durations, and acoustic waveform data pieces of adjacent segments may have an overlap of five minutes. In this case, the subsequent data processing is executed on a segment-by-segment basis.
- feature amount extraction section 110 applies a high-pass filter to the acoustic data to remove a low-frequency noise.
- feature amount extraction section 110 removes the noise by utilizing a mean and a variance of the acoustic data expressed by following Expression 1 and the Gaussian model:
- feature amount extraction section 110 computes, in accordance with following Expression 2, probability p(x i ) of occurrence of acoustic data x i :
- feature amount extraction section 110 performs feature amount extraction on the segments obtained by the segmentation after the preprocessing. Specifically, feature amount extraction section 110 extracts, from the segments, a temporal feature amount or statistic including mean, median, standard deviation, maximum value/minimum value, kurtosis, skewness, zero crossing rate, energy, energetic entropy, and the like.
- the extracted feature amount vector is expressed by following Expression 3:
- d 1 denotes the total number of feature amounts and “ ⁇ 1” denotes a one-dimensional vector.
- Feature amount extraction section 110 also extracts, from the segments, a time-frequency feature amount indicating a local variation pattern of the water flowing sound.
- the time-frequency feature amount includes mel-frequency cepstral coefficient (MFCC), gammatone cepstral coefficient (GTCC), mel-scale spectrogram, and the like.
- the extracted feature amount vector is expressed by following Expression 4:
- d 2 denotes the total number of the above feature amounts and “ ⁇ 1” denotes a one-dimensional vector.
- Feature amount extraction section 110 further extracts, from the segments, a feature amount for explaining the complexity of the acoustic pattern, which includes spectral center, spectral spread, spectral entropy, spectral flux, spectral roll-off, and the like.
- the extracted feature amount vector is expressed by following Expression 5:
- d 3 denotes the total number of the above feature amounts and “ ⁇ 1” denotes a one-dimensional vector.
- Feature amount extraction section 110 can obtain a feature amount vector of each of the segments expressed by following Expression 6 by concatenating the respective feature amount vectors obtained by the above-described feature amount extraction:
- feature amount extraction section 110 further provides the obtained feature amount vectors with subscripts regarding date and time and rearranges the feature amount vectors.
- feature amount extraction section 110 obtains, for each of the segmented acoustic data pieces, the feature amount vector expressed by following Expression 7, whose elements are composed of D feature amounts:
- the feature amount vectors corresponding to the respective segments are expressed with subscripts 1 to L being given in chronological order.
- an anomaly can be searched for from the below-described two aspects.
- the acoustic pattern of the sound of flowing water including the unknown water exhibits an anomaly in terms of a short-time spectral distribution, as compared to the acoustic pattern of the normally used water.
- the time of occurrence of the unknown water is irregular. For example, the unknown water occurs at midnight; such an irregularity is hardly observed in the acoustic pattern of the normally used water.
- the obtained feature amounts are rearranged in a matrix format as illustrated in FIG. 4 in accordance with the above-mentioned subscripts. This makes it possible to emphasize time dependence of D acoustic feature amounts.
- the rearranged feature amounts are represented as f d,t .
- the format expressing the acoustic feature amount pattern generated by feature amount extraction section 110 is not limited to such a matrix format, and may also be chronological data of any other appropriate acoustic feature amounts that can express the acoustic feature amount pattern.
- Unknown-water prediction section 120 predicts the presence or absence of the unknown water from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from the data including one or both of the acoustic data including the water flowing sound during no-rainfall time and the acoustic data of the water flowing sound under a condition different from the no-rainfall time.
- unknown-water prediction section 120 has a learned unknown-water detection model into which the acoustic feature amount pattern is inputted and that outputs a prediction result on the presence or absence of the unknown water, and determines, by utilizing the learned unknown-water detection model, whether the target acoustic data for prediction indicates the presence of the unknown water, based on the acoustic feature amount pattern of the target acoustic data for prediction generated by feature amount extraction section 110 .
- the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data obtained during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge.
- the unknown water typically arises from rainwater during rainfall or groundwater
- the acoustic data during the rainfall time can be used as the acoustic data including the water flowing sound of unknown water.
- it is possible to obtain the acoustic data including the water flowing sound of unknown water by artificially adding flowing water to the flowing water flowing through a sewer pipe.
- the unknown-water detection model may utilize a subspace as determined in accordance with a subspace method, which indicates the acoustic feature amount pattern of acoustic data including the water flowing sound during no-rainfall time that does not include the water flowing sound of unknown water.
- the unknown-water detection model may be based on a neural network that distinguishes between the acoustic feature amount pattern of acoustic data including the sound of flowing water including unknown water under a condition different from the no-rainfall time, on the one hand, and the acoustic feature amount pattern of acoustic data including the water flowing sound during the no-rainfall time that does not include the water flowing sound of unknown water, on the other hand.
- FIG. 5 is a flowchart illustrating an unknown-water detection process according to one example of the present invention.
- the unknown-water detection process is executed by unknown-water detection apparatus 100 , more specifically, by processor 101 of unknown-water detection apparatus 100 .
- unknown-water detection apparatus 100 preprocesses the acoustic data. Specifically, as described above, unknown-water detection apparatus 100 performs various preprocessing and segmentation, such as removal of noise included in the obtained acoustic data and segmentation of the acoustic data into data pieces of a predetermined time duration.
- unknown-water detection apparatus 100 extracts a feature amount(s) from the preprocessed acoustic data.
- the feature amount to be extracted may be a temporal feature amount or statistic (e.g., mean, median, standard deviation, maximum value/minimum value, kurtosis, skewness, zero crossing rate, energy, energetic entropy, and the like), mel-frequency cepstral coefficient (MFCC), gammatone cepstral coefficient (GTCC), time-frequency feature amount (e.g., mel-scale spectrogram, or the like), and other feature amounts (e.g., spectral center, spectral spread, spectral entropy, spectral flux, spectral roll-off, and the like), but are not limited thereto.
- MFCC mel-frequency cepstral coefficient
- GTCC gammatone cepstral coefficient
- time-frequency feature amount e.g., mel-scale spectrogram, or the like
- other feature amounts e
- unknown-water detection apparatus 100 generates the acoustic feature amount pattern from the extracted feature amount(s). Specifically, as described above, unknown-water detection apparatus 100 may generate any suitable chronological data of the acoustic feature amounts (e.g., acoustic feature amount pattern in a matrix format as illustrated in FIG. 4 ) that can express the acoustic feature amount pattern of each of the various acoustic feature amounts.
- acoustic feature amount pattern e.g., acoustic feature amount pattern in a matrix format as illustrated in FIG. 4
- unknown-water detection apparatus 100 may generate a one-week acoustic feature amount pattern for each of the feature amounts extracted at step S 102 (143 types of feature amounts are extracted in the illustrated specific example).
- the one-day acoustic feature amount patterns tend to have substantially constant patterns.
- a one-day water supply usage has a pattern as illustrated in FIG. 7 A . The water supply is hardly used at midnight, and the water supply usage tends to increase in the time slots before and after breakfast and dinner when a family is mainly active at home.
- the acoustic feature amount pattern of the water flowing sound also has generally the same pattern correspondingly although there is a time lag between the time of water supply usage and the time after drainage to a sewer pipe as illustrated in FIG. 7 B .
- a deviation from the illustrated acoustic feature amount pattern arises. It is thus possible to detect the unknown water by identifying such a deviated acoustic feature amount pattern.
- unknown-water detection apparatus 100 utilizes a learned machine learning model to predict the presence or absence of unknown water from the acoustic feature amount pattern. More specifically, as described above, unknown-water detection apparatus 100 has the learned unknown-water detection model into which the acoustic feature amount pattern is inputted and that outputs a prediction result on the presence or absence of unknown water, and determines, by utilizing the learned unknown-water detection model, whether target acoustic data for prediction indicates the existence of unknown water, based on the generated acoustic feature amount pattern. For example, the prediction result on the presence or absence of unknown water may be obtained in a data format as illustrated in FIG. 8 .
- a time slot in which unknown water is suspected e.g., a time slot in which there is a significant deviation from a normal acoustic feature amount pattern not including the water flowing sound of unknown water
- the degree of possibility of occurrence of unknown water may be indicated in accordance with the degree of deviation as illustrated in the figure.
- unknown-water prediction section 120 utilizes an unknown-water detection model in which a subspace is used to characterize an acoustic feature amount pattern extracted from acoustic data during no-rainfall time that does not include the water flowing sound of unknown water.
- unknown-water prediction section 120 predicts the presence or absence of unknown water in target acoustic data for prediction, based on a score of anomaly degree that indicates a deviation from the subspace formed by the acoustic feature amount pattern extracted from acoustic data including the water flowing sound during no-rainfall time that does not include the water flowing sound of unknown water.
- unknown-water prediction section 120 has in advance information about the subspace derived in accordance with the subspace method from the acoustic feature amount pattern extracted from the acoustic data indicating the normally used water during no-rainfall time that does not include the water flowing sound of unknown water, computes a score of anomaly degree that indicates a deviation between an acoustic feature amount extracted from the target acoustic data for prediction and the subspace, and predicts the presence or absence of unknown water in accordance with the computed score of anomaly degree. For example, as illustrated in FIG.
- unknown-water prediction section 120 may have in advance subspace S sa (U, ⁇ ) indicating the acoustic feature amount pattern of the normally used water, compute distance d(x, S sa (U, ⁇ )) between target acoustic feature amount pattern x for prediction and subspace S sa (U, ⁇ ) as the score of anomaly degree, and determine that the inputted acoustic feature amount pattern includes unknown water when computed distance d(x, S sa (U, ⁇ )) is greater than or equal to a predetermined threshold.
- such a subspace may be derived as described below from a training acoustic feature amount pattern extracted from acoustic data including acoustic data during no-rainfall time and acoustic data during rainfall time.
- the subspace of the acoustic feature amount pattern of normally used water is derived in accordance with the subspace method from the acoustic feature amount pattern indicating the normally used water during no-rainfall time that does not include unknown water.
- Acoustic feature amount patterns including patterns for no-rainfall time and for rainfall time are divided into the acoustic feature amount pattern for the no-rainfall time and the acoustic feature amount pattern for the rainfall time. Since the unknown water mainly occurs during rainfall time but not during no-rainfall time, the subspace is derived by utilizing the acoustic feature amount pattern for the no-rainfall time.
- a D-dimensional feature amount vector extracted from acoustic data during the no-rainfall time is expressed by following Expression 8:
- a covariance matrix can be computed by following Expression 9, and eigenvalue decomposition (EVD) is performed on the covariance matrix:
- the projection operator on an orthogonal complementary space of S is expressed as following Expression 16, and a deviation (similarity) of feature amount vector f from subspace S can be computed as expressed in following Expression 17, in which “r 2 ” is used as the score of anomaly degree that indicates the deviation of feature amount vector f from subspace S:
- a feature amount vector extracted from acoustic data during no-rainfall time can be represented as a linear combination of basis vectors of the subspace, and the value of the score of anomaly degree r 2 is 0 or a value close to 0.
- a feature amount vector extracted from acoustic data including the water flowing sound of unknown water is linearly independent of the basis vectors of the subspace, and the value of the score of anomaly degree r 2 is significantly different from 0.
- r train denotes a deviation computed from the acoustic feature amount for no-rainfall time.
- FIG. 10 is a flowchart illustrating an unknown-water prediction process using the subspace method according to one example of the present invention.
- the subspace derived based on the subspace method is utilized as the unknown-water detection model at step S 104 of FIG. 5 .
- unknown-water detection apparatus 100 computes a deviation between the subspace characterizing the acoustic feature amount pattern of the normally used water derived as described above and a target feature amount vector for prediction.
- unknown-water detection apparatus 100 detects the presence or absence of unknown water based on comparison between the computed deviation and a predetermined threshold. Specifically, when the computed deviation is greater than the predetermined threshold, unknown-water detection apparatus 100 determines that unknown water exists, and when the computed deviation is less than or equal to the predetermined threshold, unknown-water detection apparatus 100 determines that unknown water does not exist.
- the threshold may be set, for example, based on the mean and standard deviation of the distribution of the deviations computed from the acoustic feature amounts for no-rainfall time in accordance with Expression 5.
- unknown-water detection apparatus 100 may also determine the strength of suspicion of unknown water in accordance with the degree of excess. For example, unknown-water detection apparatus 100 may set the strength of suspicion of the unknown water to al when the deviation is greater than the threshold and less than or equal to twice the threshold and set the strength of suspicion of the unknown water to ⁇ 2 (> ⁇ 1 ) when the deviation is greater than twice the threshold and less than or equal to three times the threshold, and so on.
- the higher the strength of suspicion i.e., the greater the deviation
- the higher the reliability of determination that unknown water is included and the lower the strength of suspicion (i.e., the less significant the deviation) the lower the reliability of determination that unknown water is included.
- sewer pipes can be roughly classified into a branch sewer pipe disposed in a position where the amount of water is comparatively small and a trunk sewer pipe in which flowing water through the branch sewer pipe gathers.
- a branch sewer pipe even a relatively small amount of inflow of unknown water may change the amount of water and also change the water flowing sound.
- the trunk sewer pipe it may be impossible to discern the change in the amount of water and the water flowing sound unless a large amount of unknown water flows into the pipe. For this reason, the threshold may be adjusted in accordance with the amount of water that normally flows through the sewer pipe.
- a threshold set for the trunk sewer pipe in which the amount of water is relatively large may be relatively smaller than a threshold set for the branch sewer pipe in which the amount of water is relatively small, such that even a relatively small change in the amount of water is detected as unknown water.
- unknown-water prediction section 120 has an unknown-water detection model implemented as a neural network that learned acoustic feature amount patterns extracted from data including acoustic data during no-rainfall time that does not include the water flowing sound of unknown water and acoustic data obtained under a condition different from the no-rainfall time that includes the water flowing sound of unknown water, and inputs an acoustic feature amount pattern of target acoustic data for prediction into the unknown-water detection model to predict the presence or absence of unknown water as an output from the unknown-water detection model.
- the unknown-water detection model is a neural network that learned by utilizing, as training data, an acoustic feature amount pattern extracted from acoustic data including the water flowing sound in a sewer pipe during no-rainfall time that does not include the water flowing sound of unknown water, and an acoustic feature amount pattern extracted from acoustic data including the sound of flowing water including unknown water.
- Unknown-water prediction section 120 inputs, into the learned unknown-water detection model, an acoustic feature amount pattern extracted from acoustic data of the water flowing sound collected in a target sewer pipe for prediction and obtains an output result indicating the presence or absence of unknown water.
- FIG. 11 is a flowchart illustrating a learning process for the unknown-water detection model implemented by the neural network according to one example of the present invention.
- the learning process is typically performed by a computation apparatus other than unknown-water detection apparatus 100 , but the performer is not limited to this, and the leaning process may also be performed by unknown-water detection apparatus 100 .
- the below-described example will be described under an assumption that the learning process is performed by any computation apparatus having the same hardware configuration illustrated in FIG. 2 as unknown-water detection apparatus 100 .
- the unknown-water detection model obtained by the learning process is made available to unknown-water detection apparatus 100 , and is utilized for the unknown-water prediction process of unknown-water prediction section 120 described above.
- the computation apparatus inputs training acoustic feature amount patterns into a target neural network for learning at step S 301 .
- the acoustic feature amount pattern extracted from acoustic data including the water flowing sound in a sewer pipe during no rainfall time that does not include the water flowing sound of unknown water and the acoustic feature amount pattern extracted from acoustic data including the sound of flowing water including unknown water are utilized as the training acoustic feature amount patterns.
- the computation apparatus compares a prediction value on the presence or absence of unknown water outputted from the neural network with labels (i.e., “including unknown water” or “not including unknown water”) of the inputted acoustic feature amount pattern.
- the computation apparatus updates a parameter(s) of the neural network based on the result of the comparison.
- the update of the parameter may be performed by back propagation or the like.
- the computation apparatus determines whether an end condition has been satisfied.
- the end condition has been satisfied (S 304 : YES)
- the learning process is ended and the ultimately obtained neural network is considered as the learned unknown-water detection model.
- the end condition has not been satisfied (S 304 : NO)
- the computation apparatus repeats steps S 301 to S 303 .
- the end condition may be that a predetermined number of training data pieces have been processed, for example.
- the unknown-water detection model according to the present invention is not limited to this example, and chronological data of a predetermined period (e.g., one day, one week, or the like) such as an acoustic waveform of acoustic data, one or more feature amounts extracted from acoustic data, or the like may be inputted into the unknown-water detection model.
- a predetermined period e.g., one day, one week, or the like
- unknown-water prediction section 120 may narrow down the location of occurrence of unknown water based on acoustic data obtained at a plurality of locations. Specifically, when obtaining acoustic data from a plurality of sound collection locations (manholes), unknown-water prediction section 120 predicts the presence or absence of unknown water from acoustic feature amount patterns extracted from the respective acoustic data.
- unknown-water detection apparatus 100 described above together with one or more sound collection apparatuses, may form an unknown-water detection system.
- unknown-water detection apparatus 100 may collect acoustic data including the water flowing sound from each sound collection apparatus via a communication line(s) using one or both of radio and wire, extract an acoustic feature amount pattern from the respective acoustic data collected, and perform prediction on the presence or absence of unknown water for each of the extracted acoustic feature amount patterns by utilizing the learned unknown-water detection model as described above, for example.
- FIG. 12 illustrates an unknown-water prediction result by unknown-water detection apparatus 100 according to one example of the present invention.
- the graph on the left side of FIG. 12 represents the amount of rainfall in a certain area for one month from May 1 to May 31, and the graph on the right side illustrates an unknown-water prediction result by unknown-water detection apparatus 100 .
- Acoustic data including the sound of flowing water flowing through the target sewer pipe that had been collected over a period of one month from May 1 to May 31 was provided to unknown-water detection apparatus 100 .
- the illustrated graph indicating the prediction result the unknown water corresponding to rainfall on May 13 and 24 was detected in the sewer pipe after a time lag.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Multimedia (AREA)
- Acoustics & Sound (AREA)
- Human Computer Interaction (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Signal Processing (AREA)
- Hydrology & Water Resources (AREA)
- Fluid Mechanics (AREA)
- Electromagnetism (AREA)
- Public Health (AREA)
- Water Supply & Treatment (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Examining Or Testing Airtightness (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
- Sewage (AREA)
Abstract
Provided is a technology for realizing detection of unidentified water without using a flowmeter. One embodiment of the present invention pertains to an unidentified water detection device having: a feature quantity extraction unit which extracts, from acoustic data including running water sound, an acoustic feature quantity pattern that indicates temporal changes in an acoustic feature quantity; and an unidentified water prediction unit which, using a machine learning model learned from the acoustic feature quantity pattern extracted from data that contains acoustic data including running water sound during lack of rainfall and/or acoustic data of running water sound under a condition different from during lack of rainfall, predicts the presence/absence of unidentified water from the acoustic feature quantity pattern of acoustic data of a prediction object, wherein the acoustic feature quantity pattern indicates temporal changes in the acoustic feature quantity in terms of time period and day of the week.
Description
- The present application is a continuation application of U.S. patent application Ser. No. 17/046,351, filed on Oct. 9, 2020, the entire contents of which are incorporated herein by reference and priority to which is hereby claimed. Application Ser. No. 17/046,351 is the U.S. national stage of application No. PCT/JP2019/042856 filed Oct. 31, 2019. Priority under 35 U.S.C. § 119(a) and 35 U.S.C. § 365(b) is hereby claimed from Japanese Application No. 2018-207773 filed Nov. 2, 2018, the disclosure of which are both incorporated herein by reference.
- The present invention relates to a technique for detecting unknown water using a machine learning model.
- A separate sewer system is currently in the mainstream of sewage treatment, in which a wastewater pipeline for wastewater discharged from houses, factories, and the like and a rainwater pipeline for rainwater are separately installed. The wastewater discharged into the wastewater pipeline is treated in a sewage treatment plant and is let out into a river or the like after treatment. Meanwhile, the rainwater that has flowed into the rainwater pipeline is directly let out into a river or the like.
- The amount of wastewater flowing into the sewage treatment plant is regarded as a wastewater treatment amount for treatment at the sewage treatment plant. In principle, a sewerage fee corresponding to water supply usage is charged for the wastewater treatment amount (revenue water amount). However, there is actually a large difference between the wastewater treatment amount and the revenue water amount. The amount of water corresponding to this difference is called unknown water. Typically, the unknown water is rainwater, groundwater, and/or the like that somehow flow into the wastewater pipeline. The wastewater having flowed into the wastewater pipeline has to be treated in the sewage treatment plant. Thus, the unknown water places stress on management of sewage treatment business. For example, the 2011 statistics indicate that the unknown water is 17.1% of the annual total amount of treated water.
- See, for example, Japanese Patent Application Laid-Open No. 2011-080347.
- As is understood, the unknown water arising from rainwater and/or groundwater often occurs due to the inflow of rainwater during rainfall. For this reason, conventionally, the occurrence of unknown water has been forecast by using flowmeters installed at multiple sewer pipe installation locations for detecting where the flow rate in the sewer pipe increases during rainfall or after rainfall.
- However, the installation of each of the flowmeters is costly in terms of not only the flowmeter itself, but also the work cost and labor cost for installation and/or the like, and therefore, a technique for detecting unknown water which can be implemented at lower cost has been required.
- An object of the present invention considering the above-mentioned problem is to provide a technique for achieving detection of unknown water without the need for a flowmeter.
- One aspect of the present invention relates to an unknown-water detection apparatus that, in order to solve the above problem, includes a feature amount extraction section that extracts, from acoustic data including a water flowing sound, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount; and an unknown-water prediction section that predicts presence or absence of unknown water from an acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned an acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time, in which the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot.
- According to the present invention, detection of unknown water can be achieved without the need for a flowmeter.
-
FIG. 1 illustrates an outline of an unknown-water detection apparatus according to one example of the present invention; -
FIG. 2 is a block diagram illustrating a hardware configuration of the unknown-water detection apparatus according to one example of the present invention; -
FIG. 3 is a block diagram illustrating a functional configuration of the unknown-water detection apparatus according to one example of the present invention; -
FIG. 4 illustrates an acoustic feature amount pattern in a matrix format according to one example of the present invention; -
FIG. 5 is a flowchart illustrating an unknown-water detection process according to one example of the present invention; -
FIG. 6A schematically illustrates an acoustic feature amount pattern according to one example of the present invention; -
FIG. 6B schematically illustrates an acoustic feature amount pattern according to one example of the present invention; -
FIGS. 7A and 7B illustrate a specific example of a temporal change in water usage and acoustic feature amount in one day; -
FIG. 8 illustrates a prediction result on the presence or absence of unknown water according to one example of the present invention; -
FIG. 9 schematically illustrates an unknown-water detection model using a subspace method according to one example of the present invention; -
FIG. 10 is a flowchart illustrating an unknown-water prediction process using the subspace method according to one example of the present invention; -
FIG. 11 is a flowchart illustrating a learning process for an unknown-water detection model implemented using a neural network according to one example of the present invention; and -
FIG. 12 illustrates an unknown-water prediction result by the unknown-water detection apparatus according to one example of the present invention. - An embodiment of the present invention will be described below with reference to the drawings.
- In the below-mentioned example, an unknown-water detection apparatus is disclosed that detects unknown water from acoustic data including a water flowing sound in a sewer pipe. The example described below is outlined as follows: when the unknown-water detection apparatus obtains acoustic data including a water flowing sound in a sewer pipe collected by a sound collection apparatus installed down in a manhole, the unknown-water detection apparatus extracts an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount from the obtained acoustic data, and predicts the presence or absence of unknown water from the extracted acoustic feature amount pattern by utilizing a learned unknown-water detection model.
- Here, the unknown-water detection model may be a machine learning model that learned an acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time. For example, the unknown-water detection model may be a machine learning model that uses a subspace to characterize the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time (i.e., an acoustic feature amount pattern of normally used water not including unknown water). Alternatively, the unknown-water detection model may be a neural network that learned the acoustic feature amount pattern extracted from the acoustic data during the no-rainfall time that does not include the water flowing sound of unknown water, and from the acoustic data obtained under the condition different from the no-rainfall time that includes the water flowing sound of unknown water.
- Thus, the unknown-water detection apparatus can detect the presence or absence of unknown water based on the acoustic data obtained by utilizing a sound collection apparatus such as a commercially available voice recorder, so that it becomes possible to perform unknown-water detection at lower cost than in unknown-water detection using the conventional flowmeter.
- To begin with, the unknown-water detection apparatus according to one example of the present invention will be described with reference to
FIGS. 1 to 8 .FIG. 1 schematically illustrates the unknown-water detection apparatus according to one example of the present invention. - As illustrated in
FIG. 1 , when unknown-water detection apparatus 100 obtains acoustic data including a water flowing sound in a sewer pipe obtained by a sound collection apparatus such as a voice recorder installed in a manhole, the unknown-water detection apparatus utilizes an unknown-water detection model implemented as a machine learning model to predict the presence or absence of unknown water from the obtained acoustic data. - For example, the sound collection apparatus may be a commercially available voice recorder or the like. Sound collection apparatuses are installed down in a plurality of manholes in an area, and record, as acoustic data in a predetermined format or acoustic waveform, sounds (water flowing sounds) of water flowing through sewer pipes down in the manholes. The acoustic data recorded in each of the sound collection apparatuses may be appropriately collected by hands or the like or may also be collected via a communication network when the sound collection apparatus is provided with a communication function. Unknown-
water detection apparatus 100 preprocesses the obtained acoustic data and segments the obtained acoustic data into segments of data of a predetermined period of time (e.g., 10 minutes) to generate an acoustic feature amount pattern indicating a temporal change of time-serialized acoustic data (e.g., in one day, one week, or the like). Unknown-water detection apparatus 100 inputs the generated acoustic feature amount pattern into the learned unknown-water detection model, to obtain a prediction result indicating the presence or absence of unknown water from the unknown-water detection model. - Here, unknown-
water detection apparatus 100 may have a hardwareconfiguration including processor 101 such as a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), or the like,memory 102 such as a Random Access Memory (RAM), a flash memory, or the like,hard disk 103, and input/output (I/O)interface 104 as illustrated inFIG. 2 , for example. -
Processor 101 executes various processes of unknown-water detection apparatus 100, which will be described in detail below. -
Memory 102 stores therein various data and programs for unknown-water detection apparatus 100 together with a program for implementing the unknown-water detection model, and functions as a working memory especially for working data, a program being executed, and the like. Specifically,memory 102 stores therein data and programs loaded fromhard disk 103 and functions as the working memory during program execution byprocessor 101. -
Hard disk 103 stores therein the various data and programs for unknown-water detection apparatus 100 together with the program for implementing the unknown-water detection model. - I/
O interface 104 is an interface for inputting and outputting data into and from an external apparatus, and is a device for inputting and outputting data of a Universal Serial Bus (USB), a communication circuit, a keyboard, a mouse, a display, and the like, for example. - However, the hardware configuration of unknown-
water detection apparatus 100 according to the present disclosure is not limited to the above-described configuration, and the unknown-water detection apparatus may have any other suitable hardware configuration. For example, a part or all of the unknown-water detection process of unknown-water detection apparatus 100 described above may be implemented by a processing circuit or an electronic circuit wired to implement the unknown-water detection process. -
FIG. 3 is a block diagram illustrating a functional configuration of unknown-water detection apparatus 100 according to one example of the present invention. As illustrated inFIG. 3 , unknown-water detection apparatus 100 includes featureamount extraction section 110 and unknown-water prediction section 120. - Feature
amount extraction section 110 extracts, from the acoustic data including the water flowing sound, the acoustic feature amount pattern indicating a temporal change in an acoustic feature amount. - Specifically, feature
amount extraction section 110 first performs preprocessing and segmentation on the acoustic data indicating an acoustic waveform representing the water flowing sound in a sewer pipe. When obtained the acoustic data, featureamount extraction section 110 first discards, when the acoustic data includes any data loss caused by a malfunction of the sound collection apparatus or the like, a zero value or non-numerical data in the lost part. Then, featureamount extraction section 110 segments the obtained continuous acoustic waveform into segments at predetermined time intervals in order to facilitate subsequent data processing. The segmentation may divide the acoustic waveform data into data pieces having 10-minute durations, and acoustic waveform data pieces of adjacent segments may have an overlap of five minutes. In this case, the subsequent data processing is executed on a segment-by-segment basis. - Further, in order to remove or reduce an interference signal indicating an ambient noise (e.g., vehicle horn, noise generated by a tire passing over a manhole cover, or the like) that gets mixed in the acoustic data during collection of the acoustic data, feature
amount extraction section 110 applies a high-pass filter to the acoustic data to remove a low-frequency noise. The acoustic data after the application of the high-pass filter is expressed as x=[x1, . . . , xN]. In this expression, “N” denotes the number of acoustic data pieces in one segment. - Thereafter, feature
amount extraction section 110 removes the noise by utilizing a mean and a variance of the acoustic data expressed by followingExpression 1 and the Gaussian model: -
- Specifically, when acoustic data string x in a segment is given, feature
amount extraction section 110 computes, in accordance with followingExpression 2, probability p(xi) of occurrence of acoustic data xi: -
- When p(xi)<ε(where ε is a predetermined threshold for noise removal), inputted segment xi is determined as an anomalous state or anomaly and is deleted from target acoustic data to be processed.
- Next, feature
amount extraction section 110 performs feature amount extraction on the segments obtained by the segmentation after the preprocessing. Specifically, featureamount extraction section 110 extracts, from the segments, a temporal feature amount or statistic including mean, median, standard deviation, maximum value/minimum value, kurtosis, skewness, zero crossing rate, energy, energetic entropy, and the like. The extracted feature amount vector is expressed by following Expression 3: -
(3) - In this expression, “d1” denotes the total number of feature amounts and “×1” denotes a one-dimensional vector.
- Feature
amount extraction section 110 also extracts, from the segments, a time-frequency feature amount indicating a local variation pattern of the water flowing sound. The time-frequency feature amount includes mel-frequency cepstral coefficient (MFCC), gammatone cepstral coefficient (GTCC), mel-scale spectrogram, and the like. - The extracted feature amount vector is expressed by following Expression 4:
-
(4) - In this expression, “d2” denotes the total number of the above feature amounts and “×1” denotes a one-dimensional vector.
- Feature
amount extraction section 110 further extracts, from the segments, a feature amount for explaining the complexity of the acoustic pattern, which includes spectral center, spectral spread, spectral entropy, spectral flux, spectral roll-off, and the like. The extracted feature amount vector is expressed by following Expression 5: -
(5) - In this expression, “d3” denotes the total number of the above feature amounts and “×1” denotes a one-dimensional vector.
- Feature
amount extraction section 110 can obtain a feature amount vector of each of the segments expressed by followingExpression 6 by concatenating the respective feature amount vectors obtained by the above-described feature amount extraction: -
(6) - In this expression, D=d1+d2+d3.
- Next, feature
amount extraction section 110 further provides the obtained feature amount vectors with subscripts regarding date and time and rearranges the feature amount vectors. As described above, featureamount extraction section 110 obtains, for each of the segmented acoustic data pieces, the feature amount vector expressed by followingExpression 7, whose elements are composed of D feature amounts: -
(7) - In the above expression, the feature amount vectors corresponding to the respective segments are expressed with
subscripts 1 to L being given in chronological order. For example, in order to detect unknown water, an anomaly can be searched for from the below-described two aspects. In respect of the first aspect, the acoustic pattern of the sound of flowing water including the unknown water exhibits an anomaly in terms of a short-time spectral distribution, as compared to the acoustic pattern of the normally used water. In respect of the second aspect, the time of occurrence of the unknown water is irregular. For example, the unknown water occurs at midnight; such an irregularity is hardly observed in the acoustic pattern of the normally used water. In order to effectively search for the anomaly from these two aspects, the obtained feature amounts are rearranged in a matrix format as illustrated inFIG. 4 in accordance with the above-mentioned subscripts. This makes it possible to emphasize time dependence of D acoustic feature amounts. The rearranged feature amounts are represented as fd,t. - However, the format expressing the acoustic feature amount pattern generated by feature
amount extraction section 110 is not limited to such a matrix format, and may also be chronological data of any other appropriate acoustic feature amounts that can express the acoustic feature amount pattern. - Unknown-
water prediction section 120 predicts the presence or absence of the unknown water from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from the data including one or both of the acoustic data including the water flowing sound during no-rainfall time and the acoustic data of the water flowing sound under a condition different from the no-rainfall time. Specifically, unknown-water prediction section 120 has a learned unknown-water detection model into which the acoustic feature amount pattern is inputted and that outputs a prediction result on the presence or absence of the unknown water, and determines, by utilizing the learned unknown-water detection model, whether the target acoustic data for prediction indicates the presence of the unknown water, based on the acoustic feature amount pattern of the target acoustic data for prediction generated by featureamount extraction section 110. - Here, the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data obtained during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge. As described above, it is known that the unknown water typically arises from rainwater during rainfall or groundwater, and the acoustic data during the rainfall time can be used as the acoustic data including the water flowing sound of unknown water. Alternatively, it is possible to obtain the acoustic data including the water flowing sound of unknown water by artificially adding flowing water to the flowing water flowing through a sewer pipe.
- As will be described in detail below, the unknown-water detection model may utilize a subspace as determined in accordance with a subspace method, which indicates the acoustic feature amount pattern of acoustic data including the water flowing sound during no-rainfall time that does not include the water flowing sound of unknown water. Alternatively, the unknown-water detection model may be based on a neural network that distinguishes between the acoustic feature amount pattern of acoustic data including the sound of flowing water including unknown water under a condition different from the no-rainfall time, on the one hand, and the acoustic feature amount pattern of acoustic data including the water flowing sound during the no-rainfall time that does not include the water flowing sound of unknown water, on the other hand.
-
FIG. 5 is a flowchart illustrating an unknown-water detection process according to one example of the present invention. The unknown-water detection process is executed by unknown-water detection apparatus 100, more specifically, byprocessor 101 of unknown-water detection apparatus 100. - As illustrated in
FIG. 5 , at step S101, unknown-water detection apparatus 100 preprocesses the acoustic data. Specifically, as described above, unknown-water detection apparatus 100 performs various preprocessing and segmentation, such as removal of noise included in the obtained acoustic data and segmentation of the acoustic data into data pieces of a predetermined time duration. - At step S102, unknown-
water detection apparatus 100 extracts a feature amount(s) from the preprocessed acoustic data. Specifically, as described above, the feature amount to be extracted may be a temporal feature amount or statistic (e.g., mean, median, standard deviation, maximum value/minimum value, kurtosis, skewness, zero crossing rate, energy, energetic entropy, and the like), mel-frequency cepstral coefficient (MFCC), gammatone cepstral coefficient (GTCC), time-frequency feature amount (e.g., mel-scale spectrogram, or the like), and other feature amounts (e.g., spectral center, spectral spread, spectral entropy, spectral flux, spectral roll-off, and the like), but are not limited thereto. - At step S103, unknown-
water detection apparatus 100 generates the acoustic feature amount pattern from the extracted feature amount(s). Specifically, as described above, unknown-water detection apparatus 100 may generate any suitable chronological data of the acoustic feature amounts (e.g., acoustic feature amount pattern in a matrix format as illustrated inFIG. 4 ) that can express the acoustic feature amount pattern of each of the various acoustic feature amounts. - For example, as illustrated in
FIGS. 6A and 6B , unknown-water detection apparatus 100 may generate a one-week acoustic feature amount pattern for each of the feature amounts extracted at step S102 (143 types of feature amounts are extracted in the illustrated specific example). Here, the one-day acoustic feature amount patterns tend to have substantially constant patterns. For example, a one-day water supply usage has a pattern as illustrated inFIG. 7A . The water supply is hardly used at midnight, and the water supply usage tends to increase in the time slots before and after breakfast and dinner when a family is mainly active at home. It will be understood that the acoustic feature amount pattern of the water flowing sound also has generally the same pattern correspondingly although there is a time lag between the time of water supply usage and the time after drainage to a sewer pipe as illustrated inFIG. 7B . When unknown water occurs during rainfall or the like, a deviation from the illustrated acoustic feature amount pattern arises. It is thus possible to detect the unknown water by identifying such a deviated acoustic feature amount pattern. - At step S104, unknown-
water detection apparatus 100 utilizes a learned machine learning model to predict the presence or absence of unknown water from the acoustic feature amount pattern. More specifically, as described above, unknown-water detection apparatus 100 has the learned unknown-water detection model into which the acoustic feature amount pattern is inputted and that outputs a prediction result on the presence or absence of unknown water, and determines, by utilizing the learned unknown-water detection model, whether target acoustic data for prediction indicates the existence of unknown water, based on the generated acoustic feature amount pattern. For example, the prediction result on the presence or absence of unknown water may be obtained in a data format as illustrated inFIG. 8 . That is, a time slot in which unknown water is suspected (e.g., a time slot in which there is a significant deviation from a normal acoustic feature amount pattern not including the water flowing sound of unknown water) may be identified, and the degree of possibility of occurrence of unknown water may be indicated in accordance with the degree of deviation as illustrated in the figure. - Next, the unknown-water detection model based on the subspace method according to one example of the present invention will be described with reference to
FIGS. 9 and 10 . In the unknown-water detection model based on the subspace method, unknown-water prediction section 120 utilizes an unknown-water detection model in which a subspace is used to characterize an acoustic feature amount pattern extracted from acoustic data during no-rainfall time that does not include the water flowing sound of unknown water. That is, unknown-water prediction section 120 predicts the presence or absence of unknown water in target acoustic data for prediction, based on a score of anomaly degree that indicates a deviation from the subspace formed by the acoustic feature amount pattern extracted from acoustic data including the water flowing sound during no-rainfall time that does not include the water flowing sound of unknown water. - Specifically, unknown-
water prediction section 120 has in advance information about the subspace derived in accordance with the subspace method from the acoustic feature amount pattern extracted from the acoustic data indicating the normally used water during no-rainfall time that does not include the water flowing sound of unknown water, computes a score of anomaly degree that indicates a deviation between an acoustic feature amount extracted from the target acoustic data for prediction and the subspace, and predicts the presence or absence of unknown water in accordance with the computed score of anomaly degree. For example, as illustrated inFIG. 9 , unknown-water prediction section 120 may have in advance subspace Ssa(U, μ) indicating the acoustic feature amount pattern of the normally used water, compute distance d(x, Ssa(U, μ)) between target acoustic feature amount pattern x for prediction and subspace Ssa (U, μ) as the score of anomaly degree, and determine that the inputted acoustic feature amount pattern includes unknown water when computed distance d(x, Ssa (U, μ)) is greater than or equal to a predetermined threshold. - For example, such a subspace may be derived as described below from a training acoustic feature amount pattern extracted from acoustic data including acoustic data during no-rainfall time and acoustic data during rainfall time.
- To begin with, the subspace of the acoustic feature amount pattern of normally used water is derived in accordance with the subspace method from the acoustic feature amount pattern indicating the normally used water during no-rainfall time that does not include unknown water. Acoustic feature amount patterns including patterns for no-rainfall time and for rainfall time are divided into the acoustic feature amount pattern for the no-rainfall time and the acoustic feature amount pattern for the rainfall time. Since the unknown water mainly occurs during rainfall time but not during no-rainfall time, the subspace is derived by utilizing the acoustic feature amount pattern for the no-rainfall time.
- A D-dimensional feature amount vector extracted from acoustic data during the no-rainfall time is expressed by following Expression 8:
-
(8) - In this expression, “W” denotes the number of weeks of a data collection period and “d” and “t” denote day of week and time, respectively. Since [d, t] is common, fi,d,t can thus be represented simply by “fi.”
- A covariance matrix can be computed by following Expression 9, and eigenvalue decomposition (EVD) is performed on the covariance matrix:
-
- Two Resulting Matrices
-
(10) -
R f U=UΛ (2)(Expression 10) - are composed of eigenvectors expressed by following Expression 11 and
Expression 12, in which the eigenvectors are sorted in descending order of eigenvalue λi: -
(11) -
U=[u 1 , . . . , u M] (Expression 11); -
(12) -
Λ=diag(λ1, . . . , λM) (Expression 12). - The contribution of the k-th eigenvector to each eigenvector can be computed by following Expression 13:
-
- The first k eigenvectors expressed by following Expression 14 for which the contribution of ηk is greater than 0.99 are retained as dominant feature amounts in the acoustic data of the normally used water:
-
(14) -
U k =[u 1 , . . . ,u k] (Expression 14). - A subspace of an acoustic feature amount pattern of these dominant feature amounts is denoted by S and its projection operator is expressed as following Expression 15:
-
(15) -
P s =U k U k T (Expression 15). - The projection operator on an orthogonal complementary space of S is expressed as following
Expression 16, and a deviation (similarity) of feature amount vector f from subspace S can be computed as expressed in following Expression 17, in which “r2” is used as the score of anomaly degree that indicates the deviation of feature amount vector f from subspace S: -
(16) -
P ⊥ =I M −P s (Expression 16) -
(17) -
r 2 =∥P ⊥s ·f∥=f T(I D −U k U k T)f (Expression 17). - A feature amount vector extracted from acoustic data during no-rainfall time can be represented as a linear combination of basis vectors of the subspace, and the value of the score of anomaly degree r2 is 0 or a value close to 0. On the other hand, a feature amount vector extracted from acoustic data including the water flowing sound of unknown water is linearly independent of the basis vectors of the subspace, and the value of the score of anomaly degree r2 is significantly different from 0.
- By utilizing following Expression 18, it is determined that unknown water is included when r is greater than the value of threshold τ, and it is determined that unknown water is not included when r is less than or equal to the value of threshold τ:
-
(18) -
τ=mean(r train)+2*std(r train),i=1, . . . ,N (5)(Expression 18). - In this expression, “rtrain” denotes a deviation computed from the acoustic feature amount for no-rainfall time.
-
FIG. 10 is a flowchart illustrating an unknown-water prediction process using the subspace method according to one example of the present invention. In the unknown-water prediction process, the subspace derived based on the subspace method is utilized as the unknown-water detection model at step S104 ofFIG. 5 . - At step S201, unknown-
water detection apparatus 100 computes a deviation between the subspace characterizing the acoustic feature amount pattern of the normally used water derived as described above and a target feature amount vector for prediction. - At step S202, unknown-
water detection apparatus 100 detects the presence or absence of unknown water based on comparison between the computed deviation and a predetermined threshold. Specifically, when the computed deviation is greater than the predetermined threshold, unknown-water detection apparatus 100 determines that unknown water exists, and when the computed deviation is less than or equal to the predetermined threshold, unknown-water detection apparatus 100 determines that unknown water does not exist. Here, the threshold may be set, for example, based on the mean and standard deviation of the distribution of the deviations computed from the acoustic feature amounts for no-rainfall time in accordance withExpression 5. - Further, when the computed deviation is greater than the predetermined threshold, unknown-
water detection apparatus 100 may also determine the strength of suspicion of unknown water in accordance with the degree of excess. For example, unknown-water detection apparatus 100 may set the strength of suspicion of the unknown water to al when the deviation is greater than the threshold and less than or equal to twice the threshold and set the strength of suspicion of the unknown water to α2(>α1) when the deviation is greater than twice the threshold and less than or equal to three times the threshold, and so on. The higher the strength of suspicion (i.e., the greater the deviation), the higher the reliability of determination that unknown water is included, and the lower the strength of suspicion (i.e., the less significant the deviation), the lower the reliability of determination that unknown water is included. - Note that, sewer pipes can be roughly classified into a branch sewer pipe disposed in a position where the amount of water is comparatively small and a trunk sewer pipe in which flowing water through the branch sewer pipe gathers. For example, in the branch sewer pipe, even a relatively small amount of inflow of unknown water may change the amount of water and also change the water flowing sound. On the other hand, in the trunk sewer pipe, it may be impossible to discern the change in the amount of water and the water flowing sound unless a large amount of unknown water flows into the pipe. For this reason, the threshold may be adjusted in accordance with the amount of water that normally flows through the sewer pipe. For example, a threshold set for the trunk sewer pipe in which the amount of water is relatively large may be relatively smaller than a threshold set for the branch sewer pipe in which the amount of water is relatively small, such that even a relatively small change in the amount of water is detected as unknown water.
- Next, an unknown-water detection model based on a neural network according to one example of the present invention will be described with reference to
FIG. 11 . As for the unknown-water detection model based on the neural network, unknown-water prediction section 120 has an unknown-water detection model implemented as a neural network that learned acoustic feature amount patterns extracted from data including acoustic data during no-rainfall time that does not include the water flowing sound of unknown water and acoustic data obtained under a condition different from the no-rainfall time that includes the water flowing sound of unknown water, and inputs an acoustic feature amount pattern of target acoustic data for prediction into the unknown-water detection model to predict the presence or absence of unknown water as an output from the unknown-water detection model. Specifically, the unknown-water detection model is a neural network that learned by utilizing, as training data, an acoustic feature amount pattern extracted from acoustic data including the water flowing sound in a sewer pipe during no-rainfall time that does not include the water flowing sound of unknown water, and an acoustic feature amount pattern extracted from acoustic data including the sound of flowing water including unknown water. Unknown-water prediction section 120 inputs, into the learned unknown-water detection model, an acoustic feature amount pattern extracted from acoustic data of the water flowing sound collected in a target sewer pipe for prediction and obtains an output result indicating the presence or absence of unknown water. -
FIG. 11 is a flowchart illustrating a learning process for the unknown-water detection model implemented by the neural network according to one example of the present invention. The learning process is typically performed by a computation apparatus other than unknown-water detection apparatus 100, but the performer is not limited to this, and the leaning process may also be performed by unknown-water detection apparatus 100. The below-described example will be described under an assumption that the learning process is performed by any computation apparatus having the same hardware configuration illustrated inFIG. 2 as unknown-water detection apparatus 100. The unknown-water detection model obtained by the learning process is made available to unknown-water detection apparatus 100, and is utilized for the unknown-water prediction process of unknown-water prediction section 120 described above. - As illustrated in
FIG. 11 , the computation apparatus inputs training acoustic feature amount patterns into a target neural network for learning at step S301. The acoustic feature amount pattern extracted from acoustic data including the water flowing sound in a sewer pipe during no rainfall time that does not include the water flowing sound of unknown water and the acoustic feature amount pattern extracted from acoustic data including the sound of flowing water including unknown water are utilized as the training acoustic feature amount patterns. - At step S302, the computation apparatus compares a prediction value on the presence or absence of unknown water outputted from the neural network with labels (i.e., “including unknown water” or “not including unknown water”) of the inputted acoustic feature amount pattern.
- At step S303, the computation apparatus updates a parameter(s) of the neural network based on the result of the comparison. For example, the update of the parameter may be performed by back propagation or the like.
- At step S304, the computation apparatus determines whether an end condition has been satisfied. When the end condition has been satisfied (S304: YES), the learning process is ended and the ultimately obtained neural network is considered as the learned unknown-water detection model. On the other hand, when the end condition has not been satisfied (S304: NO), the computation apparatus repeats steps S301 to S303. Here, the end condition may be that a predetermined number of training data pieces have been processed, for example.
- Note that, although the acoustic feature amount patterns are inputted into the unknown-water detection model according to the present example, the unknown-water detection model according to the present invention is not limited to this example, and chronological data of a predetermined period (e.g., one day, one week, or the like) such as an acoustic waveform of acoustic data, one or more feature amounts extracted from acoustic data, or the like may be inputted into the unknown-water detection model.
- In one example, unknown-
water prediction section 120 may narrow down the location of occurrence of unknown water based on acoustic data obtained at a plurality of locations. Specifically, when obtaining acoustic data from a plurality of sound collection locations (manholes), unknown-water prediction section 120 predicts the presence or absence of unknown water from acoustic feature amount patterns extracted from the respective acoustic data. For example, when unknown-water detection apparatus 100 obtains K pieces of acoustic data DI (where “i” denotes a position index and 1≤i≤K), unknown-water prediction section 120 performs the unknown-water prediction process described above, and obtains prediction result Ji (Ji=0/1, and Ji=0 when unknown water does not exist and Ji=1 when unknown water exists) for each of the sound collection locations. That is, when Ji=1, unknown-water prediction section 120 predicts that unknown water occurs in the vicinity of the sound collection location corresponding to position index i. - That is, unknown-
water detection apparatus 100 described above, together with one or more sound collection apparatuses, may form an unknown-water detection system. Specifically, in such an unknown-water detection system, unknown-water detection apparatus 100 may collect acoustic data including the water flowing sound from each sound collection apparatus via a communication line(s) using one or both of radio and wire, extract an acoustic feature amount pattern from the respective acoustic data collected, and perform prediction on the presence or absence of unknown water for each of the extracted acoustic feature amount patterns by utilizing the learned unknown-water detection model as described above, for example. -
FIG. 12 illustrates an unknown-water prediction result by unknown-water detection apparatus 100 according to one example of the present invention. The graph on the left side ofFIG. 12 represents the amount of rainfall in a certain area for one month from May 1 to May 31, and the graph on the right side illustrates an unknown-water prediction result by unknown-water detection apparatus 100. Note that, it was known in advance that, within the area, unknown water would occur during rainfall in a sewer pipe chosen as a prediction target this time. Acoustic data including the sound of flowing water flowing through the target sewer pipe that had been collected over a period of one month from May 1 to May 31 was provided to unknown-water detection apparatus 100. As is understood from the illustrated graph indicating the prediction result, the unknown water corresponding to rainfall on May 13 and 24 was detected in the sewer pipe after a time lag. - While the examples of the present invention have been described in detail above, the present invention is not limited to the specific embodiments described above, and various modifications and changes can be made within the scope of the gist of the present invention described in the appended claims.
- This application claims the benefit of the priority of Japanese Patent Application No. 2018-207773, filed on Nov. 2, 2018, the disclosure of which including the specification, drawings and abstract is incorporated herein by reference in its entirety.
-
-
- 100 Unknown-water detection apparatus
- 110 Feature amount extraction section
- 120 Unknown-water prediction section
Claims (10)
1. An unknown-water detection apparatus, comprising:
one or more memories; and
one or more processors,
wherein the one or more processors are configured to:
extract, from acoustic data including a water flowing sound, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount; and
predict presence or absence of unknown water in a wastewater pipeline from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time,
wherein the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot, and
the machine learning model uses a subspace to characterize an acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the one or more processors are configured to predict the presence or absence of the unknown water based on a score of anomaly degree indicating a deviation from the subspace composed of the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge,
wherein the one or more processors are further configured to:
constantly receive the acoustic data over a communication network, the acoustic data collected by one or more sound collection apparatuses during an observation time period;
apply a high-pass filter to remove a low-frequency noise from the acoustic data; and
predict the presence or absence of the unknown water in a sewer pipe based on whether the deviation is larger than predetermined different thresholds corresponding to whether the sewer pipe is a brunch sewer pipe or a trunk sewer pipe.
2. The unknown-water detection apparatus according to claim 1 , wherein
the machine learning model is a neural network that learned the acoustic feature amount pattern extracted from the data including the acoustic data during the no-rainfall time and the acoustic data under the condition different from the no-rainfall time, and
the one or more processors are configured to input the acoustic feature amount pattern of the target acoustic data for prediction into the neural network to predict the presence or absence of the unknown water as an output from the neural network.
3. The unknown-water detection apparatus according to claim 1 , wherein the one or more processors are configured to narrow down a location of occurrence of the unknown water based on acoustic data obtained at a plurality of locations.
4. The unknown-water detection apparatus according to claim 1 , wherein the acoustic data is obtained via a communication line.
5. An unknown-water detection method, comprising:
extracting, by one or more processors, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount from acoustic data including a water flowing sound; and
predicting, by the one or more processors, presence, or absence of unknown water in a wastewater pipeline from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time,
wherein the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot, and
the machine learning model uses a subspace to characterize an acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the predicting includes predicting the presence or absence of the unknown water based on a score of anomaly degree indicating a deviation from the subspace composed of the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge,
wherein the method further comprises:
constantly receiving the acoustic data over a communication network, the acoustic data collected by one or more sound collection apparatuses during an observation time period;
applying a high-pass filter to remove a low-frequency noise from the acoustic data; and
predicting the presence or absence of the unknown water in a sewer pipe based on whether the deviation is larger than predetermined different thresholds corresponding to whether the sewer pipe is a brunch sewer pipe or a trunk sewer pipe.
6. The unknown-water detection method according to claim 5 , wherein
the machine learning model is a neural network that learned the acoustic feature amount pattern extracted from the data including the acoustic data during the no-rainfall time and the acoustic data under the condition different from the no-rainfall time, and
the predicting comprises inputting the acoustic feature amount pattern of the target acoustic data for prediction into the neural network to predict the presence or absence of the unknown water as an output from the neural network.
7. The unknown-water detection method according to claim 5 , wherein the predicting comprises narrowing down a location of occurrence of the unknown water based on acoustic data obtained at a plurality of locations.
8. The unknown-water detection method according to claim 5 , wherein the acoustic data is obtained via a communication line.
9. A non-transitory storage medium that stores one or more programs causing one or more processors to execute processes of:
extracting, from acoustic data including a water flowing sound, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount; and
predicting presence or absence of unknown water in a wastewater pipeline from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time,
wherein the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot, and
the machine learning model uses a subspace to characterize an acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the operation further includes predicting the presence or absence of the unknown water based on a score of anomaly degree indicating a deviation from the subspace composed of the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge,
wherein the one or more programs further causes the one or more processors to execute processes of:
constantly receiving the acoustic data over a communication network, the acoustic data collected by one or more sound collection apparatuses during an observation time period;
applying a high-pass filter to remove a low-frequency noise from the acoustic data; and
predicting the presence or absence of the unknown water in a sewer pipe based on whether the deviation is larger than predetermined different thresholds corresponding to whether the sewer pipe is a brunch sewer pipe or a trunk sewer pipe.
10. An unknown-water detection system, comprising:
one or more sound collection apparatuses; and
an unknown-water detection apparatus comprising:
one or more memories; and
one or more processors,
wherein the one or more processors are configured to:
extract, from acoustic data including a water flowing sound, an acoustic feature amount pattern indicating a temporal change in an acoustic feature amount; and
predict presence or absence of unknown water in a wastewater pipeline from the acoustic feature amount pattern of target acoustic data for prediction by utilizing a machine learning model that learned the acoustic feature amount pattern extracted from data including one or both of acoustic data including a water flowing sound during no-rainfall time and acoustic data of a water flowing sound under a condition different from the no-rainfall time,
wherein the acoustic feature amount pattern indicates a temporal change in the acoustic feature amount with respect to day of week and time slot, and
the machine learning model uses a subspace to characterize an acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the one or more processors are configured to predict the presence or absence of the unknown water based on a score of anomaly degree indicating a deviation from the subspace composed of the acoustic feature amount pattern of the acoustic data including the water flowing sound during the no-rainfall time, and
the acoustic data under the condition different from the no-rainfall time is at least one of acoustic data during rainfall time, acoustic data of a water flowing sound when another flowing water from outside is introduced into flowing water during no-rainfall time, and acoustic data of a water flowing sound due to groundwater or storm surge,
wherein the one or more processors are further configured to:
constantly receive the acoustic data over a communication network, the acoustic data collected by one or more sound collection apparatuses during an observation time period;
apply a high-pass filter to remove a low-frequency noise from the acoustic data; and
predict the presence or absence of the unknown water in a sewer pipe based on whether the deviation is larger than predetermined different thresholds corresponding to whether the sewer pipe is a brunch sewer pipe or a trunk sewer pipe.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US18/516,032 US20240086707A1 (en) | 2018-11-02 | 2023-11-21 | Device, method, program, and system for detecting unidentified water |
Applications Claiming Priority (5)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2018207773A JP6614623B1 (en) | 2018-11-02 | 2018-11-02 | Unknown water detection device, unknown water detection method, program, and unknown water detection system |
JP2018-207773 | 2018-11-02 | ||
PCT/JP2019/042856 WO2020090991A1 (en) | 2018-11-02 | 2019-10-31 | Device, method, program, and system for detecting unidentified water |
US202017046351A | 2020-10-09 | 2020-10-09 | |
US18/516,032 US20240086707A1 (en) | 2018-11-02 | 2023-11-21 | Device, method, program, and system for detecting unidentified water |
Related Parent Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2019/042856 Continuation WO2020090991A1 (en) | 2018-11-02 | 2019-10-31 | Device, method, program, and system for detecting unidentified water |
US17/046,351 Continuation US20210256372A1 (en) | 2018-11-02 | 2019-10-31 | Device, method, program, and system for detecting unidentified water |
Publications (1)
Publication Number | Publication Date |
---|---|
US20240086707A1 true US20240086707A1 (en) | 2024-03-14 |
Family
ID=68763430
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/046,351 Abandoned US20210256372A1 (en) | 2018-11-02 | 2019-10-31 | Device, method, program, and system for detecting unidentified water |
US18/516,032 Pending US20240086707A1 (en) | 2018-11-02 | 2023-11-21 | Device, method, program, and system for detecting unidentified water |
Family Applications Before (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US17/046,351 Abandoned US20210256372A1 (en) | 2018-11-02 | 2019-10-31 | Device, method, program, and system for detecting unidentified water |
Country Status (5)
Country | Link |
---|---|
US (2) | US20210256372A1 (en) |
EP (1) | EP3767556B1 (en) |
JP (1) | JP6614623B1 (en) |
AU (1) | AU2019369820B2 (en) |
WO (1) | WO2020090991A1 (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11893085B2 (en) | 2021-05-04 | 2024-02-06 | Ford Global Technologies, Llc | Anomaly detection for deep neural networks |
Family Cites Families (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS6134461A (en) * | 1984-07-26 | 1986-02-18 | Nec Corp | Equipment inspecting device |
JP3432612B2 (en) * | 1994-09-30 | 2003-08-04 | バブコック日立株式会社 | Plant operation control device |
JP2000087433A (en) * | 1998-09-11 | 2000-03-28 | Toshiba Corp | Operation control system for rainwater storage facility |
JP4399122B2 (en) * | 2001-03-27 | 2010-01-13 | メタウォーター株式会社 | Rainwater inflow prediction device |
US20020170350A1 (en) * | 2001-05-18 | 2002-11-21 | Schutzbach James S. | Method and system for analyzing the effect of inflow and infiltration on a sewer system |
CN1853098B (en) * | 2003-07-18 | 2010-12-08 | 罗斯蒙德公司 | Acoustic flowmeter and method for monitoring health degree of fixed equipment in industrial process |
FR2944903B1 (en) * | 2009-04-24 | 2016-08-26 | Thales Sa | SYSTEM AND METHOD FOR DETECTING ABNORMAL AUDIO EVENTS |
JP5574769B2 (en) | 2009-09-10 | 2014-08-20 | 株式会社東芝 | Unknown water monitoring device and unknown water monitoring method |
US9052222B2 (en) * | 2012-01-05 | 2015-06-09 | International Business Machines Corporation | Monitoring water consumption |
JP6033140B2 (en) * | 2013-03-21 | 2016-11-30 | Jxエネルギー株式会社 | Abnormality detection method and abnormality detection device |
WO2016156829A1 (en) * | 2015-03-27 | 2016-10-06 | Advanced Engineering Solutions Limited | Apparatus and method for the detection and monitoring of the condition of pipeline components |
WO2017046109A1 (en) * | 2015-09-18 | 2017-03-23 | Koninklijke Philips N.V. | Monitoring device for subject behavior monitoring |
US10914055B2 (en) * | 2015-09-25 | 2021-02-09 | Conservation Labs, Inc. | Fluid monitoring system |
US10527515B2 (en) * | 2016-04-21 | 2020-01-07 | Neptune Technology Group Inc. | Ultrasonic flow meter leak detection system and method |
FR3066969B1 (en) | 2017-06-02 | 2021-02-19 | Valeo Systemes Dessuyage | MOTOR-REDUCER FOR MOTOR VEHICLE WIPING SYSTEM |
-
2018
- 2018-11-02 JP JP2018207773A patent/JP6614623B1/en active Active
-
2019
- 2019-10-31 AU AU2019369820A patent/AU2019369820B2/en active Active
- 2019-10-31 WO PCT/JP2019/042856 patent/WO2020090991A1/en unknown
- 2019-10-31 EP EP19877630.4A patent/EP3767556B1/en active Active
- 2019-10-31 US US17/046,351 patent/US20210256372A1/en not_active Abandoned
-
2023
- 2023-11-21 US US18/516,032 patent/US20240086707A1/en active Pending
Also Published As
Publication number | Publication date |
---|---|
AU2019369820B2 (en) | 2021-04-01 |
EP3767556A4 (en) | 2021-08-04 |
EP3767556B1 (en) | 2023-03-01 |
JP2020071844A (en) | 2020-05-07 |
WO2020090991A1 (en) | 2020-05-07 |
EP3767556A1 (en) | 2021-01-20 |
US20210256372A1 (en) | 2021-08-19 |
AU2019369820A1 (en) | 2020-11-05 |
JP6614623B1 (en) | 2019-12-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110647539B (en) | Prediction method and system for vehicle faults | |
EP3836142B1 (en) | Abnormality detection device and program | |
US20240086707A1 (en) | Device, method, program, and system for detecting unidentified water | |
US20160219067A1 (en) | Method of detecting anomalies suspected of attack, based on time series statistics | |
US7991619B2 (en) | System and method using blind change detection for audio segmentation | |
CN110444011B (en) | Traffic flow peak identification method and device, electronic equipment and storage medium | |
EP3985589A1 (en) | Fraud detection device, fraud detection method, and fraud detection program | |
CN107659562A (en) | A kind of method and device for excavating malice login account | |
Zhu et al. | Spatio-temporal point processes with attention for traffic congestion event modeling | |
Akazaki et al. | Causality-aided falsification | |
CN113282920A (en) | Log abnormity detection method and device, computer equipment and storage medium | |
Dang et al. | A statistical approach for reconstructing natural streamflow series based on streamflow variation identification | |
CN113746780A (en) | Abnormal host detection method, device, medium and equipment based on host image | |
US20170242841A1 (en) | Theft detection via adaptive lexical similarity analysis of social media data streams | |
CN107783942B (en) | Abnormal behavior detection method and device | |
Kolidakis et al. | Real-time intraday traffic volume forecasting–a hybrid application using singular spectrum analysis and artificial neural networks | |
CN104282315A (en) | Voice frequency signal classified processing method, device and equipment | |
KR20170130821A (en) | Method for recommending component of software based on identifying component co-usability, and recording medium thereof | |
US20210219048A1 (en) | Acoustic signal separation apparatus, learning apparatus, method, and program thereof | |
Bai et al. | Cost-effective synchrophasor data source authentication based on multiscale adaptive coupling correlation detrended analysis | |
JP7389526B1 (en) | Optimization method for water level gauge placement, method for predicting manhole water level, and manhole water level prediction system | |
Kodag et al. | Harmonic product spectrum based approach for tonic extraction in Indian music | |
US11971332B2 (en) | Feature extraction apparatus, anomaly score estimation apparatus, methods therefor, and program | |
CN116312542A (en) | Method and device for identifying sensitive information in audio data | |
KR101925382B1 (en) | Device and Method for Water Demand Pattern Generation based on Waterworks Management DB |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: CTI ENGINEERING CO., LTD., JAPAN Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:YE, JIAXING;YOSHIDA, KEN;REEL/FRAME:065636/0844 Effective date: 20200917 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: DOCKETED NEW CASE - READY FOR EXAMINATION |